xgboost回归损失函数自定义【一】

写在前面：

每当提到损失函数，很多人都有个误解，以为用在GridSearchCV（网格搜索交叉验证“Cross Validation”）里边的scoring就是损失函数，其实并不是。我们使用构造函数构造XGBRegressor的时候，里边的objective参数才是真正的损失函数（loss function）。xgb使用sklearn api的时候需要用到的损失函数，其返回值是一阶导和二阶导，而GridSearchCV使用的scoring函数，返回的是一个float类型的数值评分（或叫准确率、或叫偏差值）。

You should be careful with the notation.

There are 2 levels of optimization here:

1. The loss function optimized when the XGBRegressor is fitted to the data.
2. The scoring function that is optimized during the grid search.

I prefer calling the second scoring function instead of loss function, since loss function usually refers to a term that is subject to optimization during the model fitting process itself.

Scikit-Learn: Custom Loss Function for GridSearchCV

因此，下文对于objective，统一叫“目标函数”；而对scoring，统一叫“评价函数”。

========== 原文分割线 ===================

许多特定的任务需要定制目标函数，来达到更优的效果。这里以xgboost的回归预测为例，介绍一下objective函数的定制过程。一个简单的例子如下：

def customObj1(real, predict):
    grad = predict - real
    hess = np.power(np.abs(grad), 0.5)
    return grad, hess

网上有许多教程定义的objective函数中的第一个参数是preds，第二个是dtrain，而本文由于使用xgboost的sklearn API，因此定制的objective函数需要与sklearn的格式相符。调用目标函数的过程如下：

model = xgb.XGBRegressor(objective=customObj1,
                         booster="gblinear")

下面是不同迭代次数的动画演示：

我们发现，不同的目标函数对模型的收敛速度影响较大，但最终收敛目标大致相同，如下图：

完整代码如下：

# coding=utf-8
import pandas as pd
import numpy as np
import xgboost as xgb
import matplotlib.pyplot as plt

plt.rcParams.update({'figure.autolayout': True})

df = pd.DataFrame({'x': [-2.1, -0.9,  0,  1,  2, 2.5,  3,  4],
                   'y': [ -10,    0, -5, 10, 20,  10, 30, 40]})
X_train = df.drop('y', axis=1)
Y_train = df['y']
X_pred = [-4, -3, -2, -1, 0, 0.4, 0.6, 1, 1.4, 1.6, 2, 3, 4, 5, 6, 7, 8]


def process_list(list_in):
    result = map(lambda x: "%8.2f" % round(float(x), 2), list_in)
    return list(result)


def customObj3(real, predict):
    grad = predict - real
    hess = np.power(np.abs(grad), 0.1)
    # print 'predict', process_list(predict.tolist()), type(predict)
    # print ' real  ', process_list(real.tolist()), type(real)
    # print ' grad  ', process_list(grad.tolist()), type(grad)
    # print ' hess  ', process_list(hess.tolist()), type(hess), '\n'
    return grad, hess


def customObj1(real, predict):
    grad = predict - real
    hess = np.power(np.abs(grad), 0.5)

    return grad, hess


for n_estimators in range(5, 600, 5):
    booster_str = "gblinear"
    model = xgb.XGBRegressor(objective=customObj1,
                             booster=booster_str,
                             n_estimators=n_estimators)
    model2 = xgb.XGBRegressor(objective="reg:linear",
                              booster=booster_str,
                              n_estimators=n_estimators)
    model3 = xgb.XGBRegressor(objective=customObj3,
                              booster=booster_str,
                              n_estimators=n_estimators)
    model.fit(X=X_train, y=Y_train)
    model2.fit(X=X_train, y=Y_train)
    model3.fit(X=X_train, y=Y_train)

    y_pred = model.predict(data=pd.DataFrame({'x': X_pred}))
    y_pred2 = model2.predict(data=pd.DataFrame({'x': X_pred}))
    y_pred3 = model3.predict(data=pd.DataFrame({'x': X_pred}))

    plt.figure(figsize=(6, 5))
    plt.axes().set(title='n_estimators='+str(n_estimators))

    plt.plot(df['x'], df['y'], marker='o', linestyle=":", label="Real Y")
    plt.plot(X_pred, y_pred, label="predict - real; |grad|**0.5")
    plt.plot(X_pred, y_pred3, label="predict - real; |grad|**0.1")
    plt.plot(X_pred, y_pred2, label="reg:linear")

    plt.xlim(-4.5, 8.5)
    plt.ylim(-25, 55)

    plt.legend()
    # plt.show()
    plt.savefig("output/n_estimators_"+str(n_estimators)+".jpg")
    plt.close()
    print(n_estimators)